Roto-translation equivariant convolutional networks: Application to histopathology image analysis

Rotation-invariance is a desired property of machine-learning models for medical image analysis and in particular computational pathology applications. We propose framework to encode the geometric structure special Euclidean motion group SE(2) convolutional networks yield translation rotation equivariance via introduction SE(2)-group convolution layers. This enables learn feature representations with discretized orientation dimension that guarantees their outputs are invariant under discrete set rotations. Conventional approaches invariance rely mostly on data augmentation, but this does not guarantee robustness output when input rotated. At that, trained conventional CNNs may require test-time augmentation reach full capability. study focused histopathology applications which it desirable arbitrary global information imaged tissues captured by machine learning models. The proposed evaluated three different tasks (mitosis detection, nuclei segmentation tumor detection). present comparative each problem show consistent increase performances can be achieved using framework.